Data and Fairness in AI

A model learns only from its training data. Balanced, representative data and testing on every group make AI fairer.

Key ideas

A model learns best what it has seen most often, so it is usually more accurate for the bigger group in its data.
Representative data includes every group in a fair share; an over-represented group appears far more than in real life.
Bias is unbalanced or unfair results for some group, often caused by one-sided training data.
Measure accuracy for each group separately to find where the model does badly.
A model that only copies past patterns can repeat past unfairness in its recommendations.
Ways to make AI fairer: add missing groups, test on different groups, and let people review decisions.

Remember

Share of a group	group count ÷ total × 100%
To balance two groups	add (larger − smaller) examples to the smaller group
Matching the largest of g groups	total afterwards = largest × g
Gap between groups	difference of the two accuracies, in percentage points

Watch out

Adding more data of the group that already has the most.
Judging fairness from one overall accuracy instead of each group’s.
Thinking a model is fair just because it is a machine.

Worked example

A dataset for a speech model has 250 samples from the east region, 150 samples from the south region, 500 samples from the north region, 100 samples from the west region. (a) How many samples are there in all? (b) What percentage come from the north region? (c) How many samples must be added to each of the other three regions, in the order listed, to match the north region? Write them in order.

Working: 500 ÷ 1000 × 100; 500 minus each

Answer: (a) 1,000; (b) 50%; (c) 250, 350, 400